Parquet supports "dictionary encoding" of column data in a manner very similar to the concept of Categoricals in pandas. It is natural to use this encoding for a column which originated as a categorical. Conversely, when loading, if the file metadata says that a given column came from a pandas (or arrow) categorical, then we can trust that the whole of the column is dictionary-encoded and load the data directly into a categorical column, rather than expanding the labels upon load and recategorising later.
If the data does not have the pandas metadata, then the guarantee cannot hold, and we cannot assume either that the whole column is dictionary encoded or that the labels are the same throughout. In this case, the current behaviour is fine.
(please forgive that some of this has already been mentioned elsewhere; this is one of the entries in the list at https://github.com/dask/fastparquet/issues/374 as a feature that is useful in fastparquet)